classifier fusion method to recognize

International Journal on Cybernetics & Informatics (IJCI) Vol. 3, No. 3, June 2014

DOI: 10.5121/ijci.2014.3301 1

CLASSIFIER FUSION METHOD TO RECOGNIZE

HANDWRITTEN PERSIAN NUMERALS

Reza Azad

1, Babak Azad

2, Iraj Mogharreb

3, Shahram Jamali

4

1Department of Electrical and Computer Engineering, Shahid Rajaee Teacher Training

University, Tehran, Iran 2

Computer Engineering Department, University of Mohaghegh Ardabili, Ardabil, Iran 3Ardabil Branch Islamic Azad University, Ardebil, Iran

4Associate professor, Faculty of Computer Engineering, University of Mohaghegh

Ardabili, Ardabil, Iran

ABSTRACT

Recognition of Persian handwritten characters has been considered as a significant field of research for

the last few years under pattern analysing technique. In this paper, a new approach for robust handwritten

Persian numerals recognition using strong feature set and a classifier fusion method is scrutinized to

increase the recognition percentage. For implementing the classifier fusion technique, we have considered

k nearest neighbour (KNN), linear classifier (LC) and support vector machine (SVM) classifiers. The

innovation of this tactic is to attain better precision with few features using classifier fusion method. For

evaluation of the proposed method we considered a Persian numerals database with 20,000 handwritten

samples. Spending 15,000 samples for training stage, we verified our technique on other 5,000 samples,

and the correct recognition ratio achievedapproximately 99.90%. Additional, we got 99.97% exactness

using four-fold cross validation procedure on 20,000 databases.

KEYWORDS

Persian handwritten recognition, k nearest neighbor, linear classifier, SVM classifier, classifier fusion.

1. INTRODUCTION

Nowadays handwritten characters recognition is one of the most popular research areas, because

it has various application potentials. Bank cheques processing, Postal Automation, Automatic

data entry, etc. are some of its potential application are. Most of the handwritten character

recognition methods for, Arabic, English, and Chinese scripts are reviewed in [1-3]. As regards

there is no popular method for Persian handwritten character recognition due to cursive-ness of

Persian handwritten, and various way of characters combination together and also, characters

position in words. By the following passages, we studied edge maps, transit and directional

frequencies effect in the numeral image contour pixels as features, which kept morphological

information of input and then applied fusion of classifiers as classifier.

In Iran and some of its neighbouring countries, the Persian numerals have usage. Also, the

Persian has 10 numerals. Alphabets of Persian and Arabic scripts are written from right to left but

in their text, digits are taken place from left to right. Despite the similarity of Persian and Arabic

numerals, there are a few important differences between their scripts [4]. Normally, in Persian


2

there are two types of writing for the digits 0, 2, 4, 5 and 6. These characteristics make Persian

numerals recognition so sophisticate than other languages [4]. Fig. 1 shows the Examples of

Persian printed and handwritten digits.

Figure 1. Sample of Persian Handwritten numerals [4]

In the related workspertinent to the Persian handwritten character recognition, many approaches

for feature extraction and classificationhas been proposed. Some of the latest methods in field of

feature extraction are; shadow and segmentation codes [5-7], fractal approach [8], profiles [9],

moment features [10], template matching [11], structural feature set [12] and wavelet [13], [14].

Also for classification stage different types of Neural Networks [5-8], [10], [11], SVM’s [9], [13],

[15], Nearest Neighbour [12], multiple classifier [16], [17] have been implemented. What

achieved from the literature survey on Persian handwritten character recognition, it is clear that

not much attempt was increased to recognize a more capable feature set (some of them are time

consuming and some of them cannot keep the structure of the input image for feature extraction

stage), which could more suitably be respond to the recognition part [4]. For solving this kind of

issue, we investigatedmore robust featureswith use of transit, edge maps and modified contour

chain code of every window-map, and then apply fusion of classifiers for classification. This kind

of feature set, which explicit the somatic shape of input character and take out local information

of the input image in each window-map, provided very good correctness in experimental stage.

We ought toremarkthat the suggested system did not use any pre-processing methods (skew and

tilt detection/improvement, smooth out, noise elimination, etc.) that were luxuriousprocesses.

Likewise, strength of our feature set was under treatment some of these issues such as skew and

slant reasonably.

The rest of our paper is organized as follows: feature extraction technique is detailed in Section

two, classification stage id described in Section three, experimental results and comparative

investigation are labelled in Section four and finally in last section we assign the conclusion.

2. FEATURE EXTRACTION TECHNIQUE

In this phase we will extract chain-code, modified edge maps and transition feature set. Extracting

tactic of these feature set are detailed in the next subsections.

2.1. Directional Chain Code FrequenciesFeature Set

Directional chain code frequencies of the outline points of the input image are very useful for

different application such as character segmentation, recognition, etc. [18]. In our proposed

method as a first feature set we extracted chain-code directional frequencies of outline pixels of

the images by the following rule: First the minimum rectangle covering the handwritten character

(bounding box) is extracted for every input image, Then for removing the features to size and


3

position, we adapted each image to a normal size of 49×49 pixels. We selected this normalized

value based of manyexperimentations and a geometric study. Fig. 2(a) shows the normalized

image with its covering rectangle and in Fig. 2(b) the extracted outline points of the character is

shown.

Figure 2. (a): bordering of a normalized image (b): Digit ‘5’ outlineshape

By possession a window-map of size 7×7 on the image, we scanned the image outline

horizontally from the top left maximum point to flatmaximum point (that contains 49 no

overlapped blocks) and we extracted a8 directions chain code frequencies for each block (8

directions were depicted in Fig. 3(a)). As a replacementextracting the feature set in terms of 8

orders, we have offered to simplify the features into 4 sets fit to 4 orders (Fig. 3(b) shows the four

directions), that the horizontal direction code are determined by the 0 and 4 directions, vertical

direction code are showed bythe direction 2 and 6, also the principal diagonal direction are

determined by the direction 1 and 5 and finallythe off diagonal direction code is masked by the 3

and 7 directions. Therefore, in every block, we acquired four values signifying the occurrences of

these four ordersand these quantities were used as local contour direction featureset.For extracting

these features, an unvarying block with 49 (7×7) size is considered in every image and we

calculated four features (four directions) in each block so we acquired 49×4=196 features for

every image.

Figure 3. (a): chain code pattern for 8 direction (b): chain code pattern for 4 direction

2.2. Modified Edge Maps Features

In this stage, at the beginning, an N x M image is converted into the thinning form and then

reshaped into a 49 × 49 matrix. For extracting the four distinct edge maps (horizontal, vertical

and two diagonals (45° and -45°)) the Sobel operators was used. After that these four distinct

edge maps are distributed into 49 sub-images of 7 × 7 pixels.Then the featureset are

gottencomputing the proportion of black pixels in each sub-image (a featureset with 49 dimension

for each image). Finally these features are joint to form a single feature vector holding 196 (49 x

4) features.

2.3. Transition features

The third feature set that investigated in this paperis based on the extracting the transition value

from background to foreground pixels in both vertical and horizontal directions. Our extracting

The transition feature set is mostly like a transition feature set that hass been proposed by Kumar

et al [19]. forextracting transition information, image is skimmed from top to bottom and left to

right. Following actions shows the way of these feature extraction.


4

Action1: Distribute the contoured image of a handwritten image into 49 part with size of 7×7.

Action2: Computethe number of transitions for each part and extract the 49 features for each

character image.

3. CLASSIFICATION BY ENSEMBLE TECHNIQUE

Ensemble technique has wonderful application in different techniques and widely used in pattern

classification and machine learning. the diversekinds of ensemble techniques has been proposed

before and among them theclassification fusion method is the most important type of the

ensemble classification. In this technique, numerous classifiers are trained on a same feature

space and then,the results of these classifiers are mixed to get a more precise classification [20].

In the current paper, we have used K-Nearest Neighbour (KNN), Linear (L) and Support Vector

Machine (SVM) classifiers for ensemble technique. The features gained from Directional Chain

Code (DCD) are applied to SVM, the Modified Edge Maps (MEM) feature Set is applied to linear

classifiers and Transition features are applied to KNN separately. The prediction of these

classifiers is combined using majority-voting procedure to appropriately classify the sample.

3.1. Classification with use of the K-Nearest Neighbour

If you faced with the classification problem has pattern classes you can use an efficient technique

called the K-nearest neighbor classifier that display a reasonably limited degree of variability.

With calculating the distance between the input pattern and the training patterns it could

recognize each input pattern with certain accuracy that has given.During the classification only k

nearest prototypes could takes into the input pattern.the final decision is performed with use of

majority of class voting.In the k-Nearest neighbour technique, the distance among train and test

set is computed for determining the class of the test set.In the Equation (1) the applying method is

detailed:

� = � (�� − ��)�� (1)

In the above equation,xk is the collection stored feature value, N shows the entire number of

features in feature set and yk is the nominee feature value.

3.2. Linear Classifier (LC)

Linear classifier (LC) is a kind of the statistical classifier thatuses a value of the linear mixture of

the featuresfor generating alabel of class.The application of the linear classifier is mostly on the

circumstances where the speed of classification is an important

issue.LCfrequentlyeffortexcellently when the number of magnitudes in feature vector is huge. It

can be signified as revealed in equation (2):

y= f(w, x) = f�∑ w�x�� (2)

Which w� is weight vector, learned from a set of marked training examples and thex� is the

feature vector of testing model and f is a simple function that applies the value to the individual

classes based on a confident threshold.


5

3.3. Support Vector Machines (SVM) Classifier

Support vector machines (SVMs) are one of the most importont, and powerfull in pattern learning

and also in pattern recognition, because it support high dimensional data and at the same time,

providing good generalization properties. Additionally, in data mining and pattern recognition

applications, SVMs have very usages. Considered that � = �(��, ��)� ! = 1are samples of n

training, where, �� ∈ �−1,1� is a class label of sample xi, and �� ∈ $%is an m-dimensional

sample in the input space. with the minimal classification errors, SVM finds the optimal

separating hyper plane (OSH). Equation (3) shown the linear separation hyper plane.

&(�) = '(� + *(3)

Here b and W are the bias, and weight vector. By solving the optimization problem (6), the

optimal hyper plane can be obtain, where variable C controls the effect of the slack variables, and

+�is slack variable for obtaining a soft margin. Decreasing the value of C cause to increasing of

separation margin. In a SVM, maximizing the generalization ability of the SVM cause to optimal

hyper plane obtained. Anyhow, the obtained classifier may not have high generalization ability in

a nonlinear separable training data. The original input space should be mapped into a high-

dimensional in order to enhancing the linear severability purpose. Now, with using the nonlinear

vector function,(�) = (,�(�), … , ,.(�))/, witch maps the m-dimensional input vector x into

the l-dimensional feature space, the OSH in the feature space is given by equation (4):

f(x) = W1φ(x) + b (4)

Equation (5) defined a decision function for a test data:

D(x) = Sign(W1φ(x) + b)(5)

By solving the following quadratic optimization problem, the optimal hyper plane can be found:

Minimize 12 ||W||� + C@ζBC

B��

SubjecttoyB(W1φ(x) + b) ≥ 1 − ζB

ζB ≥ 0, i = 1,… , n (6)

As mentioned before, SVMs classifier isintroducedfor binary problemclassification, yet, our

proposed method has more than two classes for classifications. For solving this problem,

multiclass classification strategies that mentioned in [21] can be used. The greatestcommon ones

are the one-against-one (OAO) and the one-against-all (OAA) approaches [22]. The one-against-

one isusing the (n (n-1)) ⁄2 equation for combinations of all class pairs. What achieved from the

experimental results we find out that the one-against-all is more fitting for our proposed method.

We used OAA for our character classification.

Our proposed classifier fusion method is sum up in the below algorithm. In addition Fig. 4 shows

the recognition way for entrance test image by applying fusion of classifier.


6

Algorithm Begin:

Input: a set of test samples and training samples.

Output: the class label for determining the test sample belongs.

Method: Step1: Extract the directional chain code (DCD), Modified Edge Maps (MEM) and

transit (T) features for the training sample and test sample using previously discussed approaches

respectively.

Step 2: Apply the DCD, MEM and T features obtained for the training samples to train the SVM,

KNN and L classifiers respectively and separately.

Step 3: Apply the DCD, MEM and T features obtained for the test samples to each of the

classifier.

Let the prediction of the classifiers be p1, p2 and p3

Step 4: Predict the class of the test sample as

Class = Majority of the {p1, p2, p3}

End

Figure 4. Architecture of the classifier fusion method to recognizehandwritten numerals

4. PRACTICAL RESULTS OF THE PROPOSED METHOD

For analysing of the proposed method, a set of the 15,000 samples for training stage and a set of

the 5,000 for test stage are considered as indicated in [23]. These samples contain Iranian Postal

and National Codes and were extracted with use of the 200 dpi resolution scanner from different

registration forms of the Iranian university entrance exam [24]. As we mentioned in section 2 the

writing styles of different persons, samples sizes were extremely different, we standardized them

to the constant size. By considering the 15,000 samples for the training stage, we evaluated our

method on other 5,000 samples and we achieved 99.90% precision. What achieved from the

experimental result, we obtained an exactness of 100% when the 20,000 data were utilized as

Handwritten Character Image

Extract DCD

Feature Set

Extract MEM

Feature Set

Extract T

Feature Set

0

Apply SVM

Classifier

Apply L

Classifier

Apply KNN

Classifier

0

Majority of

Vote

Recognized

Character


7

training stage and the similar dataset was used for testing stage. For further analysing we

separated our data set into 4 subclass and testing is estimated on each subclass using reminder of

the 3 subsets for training stage. The mean of the recognition rates for all the four test subclass is

achieved about 99.97%. Table 1 shows the performance comparison of the proposed method with

the state of the art methods.

Table 1. Recognition rates of the diverseapproachesfor recognizing the Persian handwritten numerals

Method

Database size Accuracy of the method (%)

Train Test Train Test

Azad et al. [4] 15000 5000 - 99.82

Shahreza et al.[5] 2600 1300 - 97.80

Harifi et al. [6] 230 500 - 97.60

Hosseini et al. [7] 480 480 - 92.00

Mozaffari et al. [8] 2240 1600 98 91.37

Rahmati et al.[9] 4979 3939 - 99.57

Dehghan et al. [10] 6000 4000 - 97.01

Faes et al. [11] 6000 4000 100 97.65

Faes et al. [12] 2240 1600 100 94.44

Mowlaei et al. [13] 2240 1600 100 92.44

Mowlaei et al. [14] 2240 1600 99.29 91.88

Sadri et al. [15] 7390 3035 - 94.14

Parvin, et al [16] 40000 2000 - 97.12

Parvin, et al [17] 60000 10000 - 98.89

Our proposed method 15000 50000 100 99.90

Our proposed method

With 4 subset

15000 50000 100 99.97

For evaluation of the proposed method we used a dataset with 20,000 samples and weachieved

99.90% and 99.97% precisions using mentioned method. In our achievement with high accuracy

(about 99.90%), we detected confusion numerals in the recognition stageamong some digits. The

main confusions were between 2, 4 and 3. This occurredsince 2, 3 and 4 resemble each other. In

Fig. 5 the success and confusion rate for each character are depicted.


Figure 5. Success and confusion rate of the proposed

5. CONCLUSIONS

In this paper, for robust Persian handwritten numerals recognition

feature extraction approachessuch as

obtaining these feature set we transformed

block based method we achieved

nearest neighbour classifiers are used for the classification. Further, the re

enhancedwith use of classifier synthesis

could be proofed that oursproposed techniquehas

numeral recognition. In addition

misclassification.Most of the misclassification

and 4, which have similar structure.

ACKNOWLEDGEMENTS

This research is supported by the S

(No.22970060-9).

REFERENCES

[1] A. Yuan, G. Bai, L. Jiao, Y. Liu, “Offline handwritten English character recognition based on

convolutional neural network “, 10th IEEE International Workshop on Document Analysis Systems,

2012, pp. 125-129.

[2] S. N. Srihari and G. Ball, "An Assessment of Arabic Handwriting Recognition Technology", Springer

book of Guide to OCR for Arabic Scripts, 2012, pp. 3

[3] F. Yin, M. Zhou, Q. Wang and C. Liu, “ Style Consistent Perturbation for Handwri

Character Recognition “, 12th IEEE International Conference onDocument Analysis and Recognition,

2013, pp. 1051-1055.

90%

91%

92%

93%

94%

95%

96%

97%

98%

99%

100%

0 1

Confusion Rate 0.05 0.03

Success Rate 99.95 99.97


and confusion rate of the proposed approachfor each Persian numeral

Persian handwritten numerals recognition we have investigated

approachessuch as directional chain code, modified edge maps and transit

transformed each image to the contour shape, then

achieved these three features sets. In the beginning, SVM, Linear and K

nearest neighbour classifiers are used for the classification. Further, the recognition accuracy was

synthesis method. What achieved from the experimental results, it

proposed techniquehas good performances on Persian handwritten

numeral recognition. In addition, in the result part we detailedthe reason of the

Most of the misclassification samples on our method were from classes of 2, 3

structure.

is supported by the Shahid Rajaee Teacher Training University, Tehr

A. Yuan, G. Bai, L. Jiao, Y. Liu, “Offline handwritten English character recognition based on


S. N. Srihari and G. Ball, "An Assessment of Arabic Handwriting Recognition Technology", Springer

book of Guide to OCR for Arabic Scripts, 2012, pp. 3-34.

F. Yin, M. Zhou, Q. Wang and C. Liu, “ Style Consistent Perturbation for Handwri


2 3 4 5 6 7 8

0.44 0.31 0.4 0.11 0.01 0 0.01

99.56 99.69 99.6 99.89 99.99 100 99.99 100


8

Persian numeral

investigated three

directional chain code, modified edge maps and transit. For

then with use of

, SVM, Linear and K

cognition accuracy was

experimental results, it

on Persian handwritten

the reason of the

were from classes of 2, 3

University, Tehran, Iran

A. Yuan, G. Bai, L. Jiao, Y. Liu, “Offline handwritten English character recognition based on


S. N. Srihari and G. Ball, "An Assessment of Arabic Handwriting Recognition Technology", Springer

F. Yin, M. Zhou, Q. Wang and C. Liu, “ Style Consistent Perturbation for Handwritten Chinese


9

0

100


9

[4] R. Azad, F. Davami and H. Shayegh, “Recognition of Handwritten Persian/Arabic Numerals Based

on Robust Feature Set and K-NN Classifier”, International Journal of Computer & Information

Technologies, Vol. 1, issue 3, 2013, pp. 220-230.

[5] M. H. Shirali-Shahreza, K. Faez and A. Khotanzad, “Recognition of Hand-written Persian/Arabic

Numerals by Shadow Coding and an Edited Probabilistic Neural Network“, IEEE International

Conference on Image Processing, Vol. 3, 1995, pp. 436-439.

[6] A. Harifi and A. Aghagolzadeh, ”A New Pattern for Handwritten Persian/Arabic Digit Recognition”,

Journal of Information Technology Vol. 3, 2004, pp. 249-252.

[7] H. Mir Mohammad Hosseini and A. Bouzerdoum, ”A Combined Method for Persian and Arabic

Handwritten Digit Recognition”, Australian New Zealand Conference on Intelligent Information

System, 1996, pp. 80 – 83.

[8] S. Mozaffari, K. Faez& H. RashidyKanan, “Recognition of Isolated Handwritten Farsi/Arabic

Alphanumeric Using Fractal Codes”, Image Analysis and Interpretation, 6th Southwest Symposium,

2004, pp. 104-108.

[9] H. Soltanzadeh and M. Rahmati, “Recognition of Persian handwritten digits using image profiles of

multiple orientations”, Pattern Recognition Letters 25, 2004, pp. 1569–1576.

[10] M. Dehghan and K. Faez, “Farsi Handwritten Character Recognition With Moment Invariants”,

Proceedings of 13th International Conference on Digital Signal Processing, Volume 2, 1997, pp. 507-

510.

[11] M. Ziaratban, K. Faez and F. Faradji, “Language-Based Feature Extraction Using Template-Matching

in Farsi/Arabic Handwritten Numeral Recognition”, Proceedings of 9th International Conference on

Document Analysis and Recognition, Vol.1, 2007, pp. 297-301.

[12] S. Mozaffari, K. Faez and M. Ziaratban, “Structural Decomposition and Statistical Description of

Farsi/Arabic Handwritten Numeric Characters”, Proceedings of the 8th Intl. Conference on Document

Analysis and Recognition, Vol. 1, 2005, pp. 237- 241.

[13] A. Mowlaei and K. Faez, “Recognition Of Isolated Handwritten Persiawarabic Characters And

Numerals Using Support Vector Machines”, Proceedings of XIII Workshop on Neural Networks for

Signal Processing, 2003, pp. 547-554.

[14] A. Mowlaei, K. Faez& A. Haghighat, ”Feature Extraction with Wavelet Transform for Recognition of

Isolated Handwritten Farsi/Arabic Characters and Numerals”, Digital Signal Processing Vol. 2, 2002,

pp. 923- 926.

[15] J. Sadri, C. Y. Suen and T. D. Bui, “Application of Support Vector Machines for Recognition of

Handwritten Arabic/Persian Digits”, Proceedings of the 2nd Conference on Machine Vision and

Image Processing & Applications, Vol. 1, 2003, pp. 300-307.

[16] H. Parvin, H. Alizadeh, M. Moshki, B. Bidgoli and N. Mozayani, ” Divide & Conquer Classification

and Optimization by Genetic Algorithm”, Third 2008 International Conference on Convergence and

Hybrid Information Technology, 2008, pp. 858 – 863.

[17] H. Parvin, H. Alizadeh, B. Bidgoli and M. Analoui, ” A Scalable Method for Improving the

Performance of Classifiers in Multiclass Applications by Pairwise Classifiers and GA”, Fourth

International Conference on Networked Computing and Advanced Information Management , 2008,

pp. 137 –142.

[18] R. Azad, B. Azad, "Real-Time Hand Gesture Recognition based on Modified Contour Chain Code

Feature Set", IJIGSP, vol.6, no.8, pp.25-31, 2014.DOI: 10.5815/ijigsp.2014.08.04.

[19] M. Kumar, M. K. Jindal and R. K. Sharma, "k -Nearest Neighbor Based Offline Handwritten

Gurmukhi Character Recognition," IEEE International Conference on Image Information Processing,

pp.1-4, 2013.

[20] D. Parikh and R. Polikar, “An Ensemble-Based Incremental Learning Approach to Data Fusion”,

IEEE Transaction on System, Man and Cybernetics, vol.37, no.2. April,2007 pp.437-450.

[21] F. Melgani and L. Bruzzone, “Classification of hyperspectral remote sensing images with support

vector machine,” IEEE Trans. Geosci. Remote Sens., vol. 42, no. 8, 2004, pp. 1778–1790.

[22] R. Azad, B. Azad and I. T. Kazeroni , “Optimized Method for Real-Time Face Recognition System

Based on PCA and Multiclass Support Vector Machine”, Advances in Computer Science: an

International Journal, Vol. 2, Issue 5, No.6 , November 2013, pp. 126-132.


10

[23] R. Azad, F. Davami and H. Shayegh, “Recognition of Handwritten Persian/Arabic Numerals Based

on Robust Feature Set and K-NN Classifier”, International Conference on computer, Information

Technology and Digital Media, 2013, pp. 161-165.

[24]H. Khosravi and E. Kabir, ”Introducing a very large dataset of handwritten Farsi digits and a study on

the variety of handwriting styles”, Pattern Recognition Letters, 28(10), 2007, pp. 1133-1141.

Authors Reza Azadobtained his B.Sc. degree with honor in computer software engineering from

SRTTU in 2014. He is IEEE & IEEE conference reviewer Member. Awarded as best

student in 2013 and 2014 by the SRTTU and awarded as best researcher in 2013 by the

SRTTU. He achieved fourth place in Iranian university entering exam. In addition he’s a

member of Iranian elites. He has a lot of scientific papers in international journal and

conferences, such as IEEE, Springer and etc. his interested research are artificial intelligence and computer

vision.

Babak Azadis a researcher from Islamic Azad University. He achieved a lot of awards

and publication on scientific papers in international journals and conferences, during his

B.Sc. education. His most interest topics are machine learning and network.

Iraj Mogharreb is a B.Sc. student in Sabalan Higher educatin institute, Ardebil, Iran, in

computer software engineering and he is top student in university. His research interests

include image processing, machine learning and information security.

ShahramJamaliis currently an Associate Professor in Mohaghegh Ardabili University,

Ardebil, Iran. He achieved his Ph.D degree in Architecture of Computer Systems in

2008 from Iran University of Science & Technology, Tehran, Iran. He has more than

100 scientific papers in international journals and conferences, such as IEEE, Elsevier,

Springer and etc. His research topics are Network security and soft computing.